Swiss-German-Speech-Data---Meeting-Notes.md

@@ -2,6 +2,31 @@
 
 ### Participants
 
+Yixuan Xu, Daniel Perruchoud, Michael Graber
+
+### Previous Action Items
+
+* [ ] Roberto: Licence agreement, also partially for Regionaljournal & Espresso broadcasts (needs to be commercially usable and openly shareable)
+* [x] FHNW : translate SFT datasets
+* [x] Prepare SRF dataset plan, share with Yixuan
+
+### Discussion Points
+
+* FHNW : \~ 3500h SRF almost ready, sanity checking ongoing,
+  * chunk size: 2050 s per batch would be possible, good would be \~ 30 - 50 s chunks
+  * global ID for original audio clip + segement should be conserved, if possible speaker ID(s)
+* Yixuan : Zürich Parliament data to be captioned with Qwen-3 omni caption, to be added to training
+  * maybe 10% of more complicated audio will be captioned like this, i.e. 50k hours
+* Roberto: first licence agreement draft available from Lars at SRF, will be shared with SVEN
+* Yixuan: audio data with translation to english, could be used for additional (80k hours, https://huggingface.co/datasets/nvidia/Granary)
+  * Would it be possible to translate Swiss German audio to english
+
+### New Action Items
+
+* [ ] FHNW: will deliver the Espresso and Regionaljournal data via huggingface dataset with download token to Yixuan
+
+### Participants
+
 Roberto Salomone, Daniel Perruchoud, Michael Graber
 
 ### Previous Action Items